CN102231203A - Image autoregressive interpolation method based on edge detection - Google Patents

Abstract

The invention discloses an image autoregressive interpolation method based on edge detection, and the method is mainly used for solving the problem of high complexity of the existing autoregressive interpolation technique. The method comprises the following steps: carrying out one quarter of downsampling on original images, carrying out edge detection on the downsampling images, approximately obtaining the edge detection images of the to-be interpolated recovery images according to the edge detection images of the downsampling images; classifying the to-be-interpolated pixels of the to-be-interpolated recovery images into to-be-interpolated pixels in smooth area and edge area, adopting bicubic to carry out interpolation on the to-be-interpolated pixels in the smooth area, adopting the autoregressive interpolation method to interpolate the to-be-interpolated pixels in the edge area; and obtaining the image after interpolation. On the premise of ensuring the visual effects of the images, the method in the invention can be utilized to carry out amplification, denoising, mending, deinterlacing and compression on the images.

Description

Based on edge-detected image autoregression interpolation method

Technical field

The invention belongs to technical field of image processing, relate to image interpolation method, this method can be used for image amplification, denoising, repairs, goes in the processing such as interlacing and compression.

Background technology

In recent years, more and more higher requirement has promoted development of digital image and widespread use to picture quality for the high speed development of computer technology and people, and method of interpolation has also obtained using widely as an important method of image processing techniques.Traditional interpolation has neighbor interpolation, bilinear interpolation bilinear and bicubic interpolation bicubic, these methods realize simple, but essence has the character of low-pass filter, and image after the interpolation is had smooth effect, has lost some important detail of the high frequency of image; Many afterwards researchists have proposed new interpolation method to keep the image border preferably, that wherein the most classical is the interpolation method NEDI based on the edge maintenance that Li and Orchard propose, this method is estimated the covariance of high-definition picture with the low-resolution image covariance, inserts the pixel value of losing according to the covariance of estimating then; In recent years, the SAI that Xiaolin Wu proposes the autoregression interpolation further improves on the NEDI basis, this method is estimated to finish by two-dimentional autoregressive modeling and piece, this algorithm comes down to a kind of adaptive inseparable two-dimentional higher order filter, SAI improves a lot on visual effect and Y-PSNR, but this method computation complexity height.

Summary of the invention

The objective of the invention is to overcome the deficiency of existing image interpolation technology, provide a kind of, reduce the computation complexity of autoregression method of interpolation based on edge-detected image autoregression interpolation method.

Realize that the above-mentioned purpose technical scheme is to add the rim detection operation in the autoregression interpolation, the interpolation pixel is divided into fringe region pixel or smooth region pixel, the fringe region pixel is used the autoregression interpolation, and the smooth region pixel is used the bicubic interpolation, and its concrete steps comprise as follows:

(1) reads in the image M that a secondary size is 512 * 512 raw form, this image is carried out 1/4th down-samplings, obtain size and be image M behind 256 * 256 the down-sampling ₁

(2) with image M behind edge detection operator Sobel and the down-sampling ₁The phase convolution obtains this edge of image image M ₂, wherein the Sobel operator has level and vertical two operators, is shown below;

$\stackrel{r}{x}=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right],$

(3) with image M behind the down-sampling ₁With edge image M ₂Carry out the recovery of 512 * 512 sizes respectively, obtain image M behind the down-sampling ₁Interpolation recover image M ' ₁With edge image M ₂Interpolation recover image M ' ₂, and with described M ₁And M ₂Each pixel is placed on M ' successively ₁And M ' ₂The position of odd-numbered line odd column is with M ' ₁And M ' ₂The pixel of all the other positions is as pixel to be inserted;

(4) to described interpolation recover image M ' ₂The known pixels of odd-numbered line odd column position averages, obtain this interpolation recover image M ' ₂The value of pixel to be inserted, the value of this pixel to be inserted is compared with predetermined threshold value T, if greater than T then this pixel to be inserted belongs to fringe region, otherwise belong to smooth region, threshold value T is between 180 to 220;

(5) described interpolation is recovered image M ' ₂Approximate as described interpolation recover image M ' ₁Edge image, with M ' ₁And M ' ₂The pixel of middle corresponding position is corresponding one by one, if M ' ₁In pixel to be inserted at smooth region, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the bicubic interpolation, obtains the value of this pixel to be inserted; If M ' ₁In pixel edge region to be inserted, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the autoregression interpolation, obtains the value of this pixel to be inserted, obtains the image M after the interpolation at last " ₁

The present invention has utilized rim detection, the pixel to be inserted of image is divided at level and smooth and fringe region, the pixel to be inserted of smooth region is used the bicubic interpolation, the bicubic interpolation is good and computation complexity is low in the interpolation of smooth region, so the present invention under the prerequisite that guarantees image visual effect after the interpolation, has reduced the computation complexity of autoregression interpolation.

Description of drawings

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the experiment simulation figure that the lena image is carried out interpolation with the present invention and existing autoregression method of interpolation;

Fig. 3 is the experiment simulation figure that the peppers image is carried out interpolation with the present invention and existing autoregression method of interpolation;

Fig. 4 is the experiment simulation figure that the boat image is carried out interpolation with the present invention and autoregression method of interpolation.

Embodiment

With reference to accompanying drawing 1, specific implementation step of the present invention is as follows:

Step 1 is read in a secondary size and is the image M of 512 * 512 raw form, and this image is carried out 1/4th down-samplings, obtains size and be image M behind 256 * 256 the down-sampling ₁

Step 2 is with image M behind edge detection operator Sobel and the down-sampling ₁The phase convolution obtains this edge of image image M ₂

The Sobel operator comprises level and vertical two operators, as (1) formula:

$\stackrel{r}{x}=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right],$

Wherein

Be horizontal operator,

Be vertical operator;

Described with image M behind edge detection operator Sobel and the down-sampling ₁The phase convolution realizes as follows:

(2a) with the horizontal component of edge detection operator Sobel

With image M behind the down-sampling ₁The phase convolution;

(2b) with the vertical component of edge detection operator Sobel

With image M behind the down-sampling ₁The phase convolution;

(2c) to (2a) and the convolution results that (2b) the obtains root of making even, obtain image M behind Sobel operator and the down-sampling ₁Convolution results.

Step 3 is with image M behind the down-sampling ₁With edge image M ₂Carry out the recovery of 512 * 512 sizes respectively, obtain image M behind the down-sampling ₁Interpolation recover image M ' ₁With edge image M ₂Interpolation recover image M ' ₂, and with described M ₁And M ₂Each pixel is placed on M ' successively ₁And M ' ₂The position of odd-numbered line odd column is with M ' ₁And M ' ₂The pixel of all the other positions is as pixel to be inserted.

Step 4, to described interpolation recover image M ' ₂The known pixels of odd-numbered line odd column position averages, obtain this interpolation recover image M ' ₂The value of pixel to be inserted, the value of this pixel to be inserted is compared with predetermined threshold value T, if the value of pixel to be inserted greater than T, then this pixel to be inserted belongs to fringe region, otherwise belongs to smooth region, threshold value T is arranged between 180 to 220.

Step 5, with described interpolation recover image M ' ₂Approximate as described interpolation recover image M ' ₁Edge image, with M ' ₁And M ' ₂The pixel of middle corresponding position is corresponding one by one, if M ' ₁In pixel to be inserted at smooth region, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the bicubic interpolation, obtains the value of this pixel to be inserted; If M ' ₁In pixel edge region to be inserted, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the autoregression interpolation, obtains the value of this pixel to be inserted, obtains the image M after the interpolation at last " ₁

Described use M ' ₁The known pixels of odd-numbered line odd column position is carried out the autoregression interpolation, realizes as follows:

(5a) use interpolation recover image M ' ₁In be in the known pixels of 5 * 5 odd-numbered line odd columns of neighborhood of pixels to be inserted position of fringe region, to M ' ₁In the known pixels of 4 * 4 odd-numbered line odd columns of this neighborhood of pixels to be inserted position carry out optimal estimation, obtain M ' ₁In the weight vector of this pixel to be inserted:

(5b) according to weight vector Utilize formula (2) calculate interpolation recover image M ' ₁In be in fringe region pixel y to be inserted ₅Value:

Wherein: y ₅Be the M ' that at every turn will insert ₁In be in the pixel to be inserted of fringe region, y ₁, y ₂, y ₃, y ₄, y ₆, y ₇, y ₈, y ₉Be M ' ₁Middle y ₅8 pixels to be inserted of neighborhood, y _iBe

In i pixel to be inserted, α _tBe In t element, x _I-tBe M ' ₁Middle y _iThe known pixels of neighborhood t odd-numbered line odd column position, x _iBe M ' ₁Middle y ₅The known pixels of 4 odd-numbered line odd columns of neighborhood position, y _I-tBe M ' ₁Middle x _i4 pixels to be inserted of neighborhood;

(5c) will calculate interpolation recover image M ' ₁In be in fringe region pixel y to be inserted ₅Value, as interpolation recover image M ' ₁In be in the autoregression interpolation of fringe region pixel to be inserted.

Effect of the present invention can further specify by following emulation experiment and contrast objective evaluation index:

(1) experiment condition:

Experimental situation of the present invention is Microsoft Visual Studio 2005, the experiment raw data is that three width of cloth sizes are the gray level image of 512 * 512 raw form, be respectively lena, peppers, boat, Fig. 2 (a) is the original image of lena, and Fig. 3 (b) is the original image of peppers, and Fig. 4 (c) is the original image of boat.

(2) experiment content:

(2.1) the lena original image shown in Fig. 2 (a) is had now the experiment simulation of autoregression method of interpolation, its simulation result is shown in Fig. 2 (b); The lena original image is carried out the experiment simulation of interpolation method of the present invention, and its simulation result is shown in Fig. 2 (c);

(2.2) the peppers original image shown in Fig. 3 (a) is had now the experiment simulation of autoregression method of interpolation, its simulation result is shown in Fig. 3 (b); The peppers original image is carried out the experiment simulation of interpolation method of the present invention, and its simulation result is shown in Fig. 3 (c);

(2.3) the boat original image shown in Fig. 4 (a) is had now the experiment simulation of autoregression method of interpolation, its simulation result is shown in Fig. 4 (b); The boat original image is carried out the experiment simulation of interpolation method of the present invention, and its simulation result is shown in Fig. 4 (c);

(2.4) interpolation method of the present invention and the autoregression interpolation method contrast simulation result on the objective evaluation index:

The objective evaluation index that the present invention uses has Y-PSNR PSNR, average gradient T and program runtime, and these objective evaluation indexs are defined as follows:

Y-PSNR PSNR: the size of supposing two width of cloth images is X * Y, make f (x, y) expression original image,

Image after the expression interpolation, objective evaluation index Y-PSNR PSNR definition as (3) formula:

$\mathrm{PSNR}=10\lg \left\{\frac{{255}^2}{\frac{1}{\mathrm{XY}}\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{[f(x,y)-\hat{f}(x,y)]}^2}\right\}---\left(3\right)$

Average gradient T: the size of supposing a sub-picture is X * Y, make f (x, the y) image of expression after the interpolation, objective evaluation index average gradient T definition as (4) formula:

$T=\frac{1}{\mathrm{XY}}\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}\sqrt{({\mathrm{\ΔI}}_x^2+\mathrm{\Δ}{I}_y^2)/2}---\left(4\right)$

△ I wherein _x=f (i+1, j)-f (i, j), △ I _y=f (i, j+1)-f (i, j);

Program runtime: whole procedure is moved the time of the usefulness that finishes.

Table 1 has provided lena, peppers, and boat three width of cloth images objective evaluation index Y-PSNR PSNR of interpolation method of the present invention and existing autoregression interpolation method, the concrete outcome of average gradient T and time, wherein black matrix is performance the best among the three.

Each evaluation index of image relatively after two kinds of method interpolation of table 1

(3) interpretation:

From the experiment simulation of interpolation method of the present invention Fig. 2 (b) as a result, Fig. 3 (b), the experiment simulation of Fig. 4 (b) and existing autoregression method of interpolation is Fig. 2 (c) as a result, Fig. 3 (c), Fig. 4 (c) more as can be seen, the visual effect of two kinds of interpolation methods of image is suitable after the interpolation.

As can be seen from Table 1: interpolation method of the present invention and existing autoregression interpolation method obtains Y-PSNR PSNR and average gradient T is suitable substantially, and the average gradient T interpolation method of the present invention of peppers and boat also will be a little more than existing autoregression interpolation method, and in the result of program runtime, interpolation method of the present invention is lower than the autoregression method of interpolation.

More than experiment shows, interpolation method of the present invention has reduced the computation complexity of autoregression interpolation under the prerequisite that guarantees image visual effect after the interpolation.

Claims (3)

1. one kind based on edge-detected image autoregression interpolation method, comprises the steps:

(1) reads in the image M that a secondary size is 512 * 512 raw form, this image is carried out 1/4th down-samplings, obtain size and be image M behind 256 * 256 the down-sampling ₁

(2) with image M behind edge detection operator Sobel and the down-sampling ₁The phase convolution obtains this edge of image image M ₂, wherein the Sobel operator has level and vertical two operators, is shown below;

$\stackrel{r}{x}=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right],$

(3) with image M behind the down-sampling ₁With edge image M ₂Carry out the recovery of 512 * 512 sizes respectively, obtain image M behind the down-sampling ₁Interpolation recover image M ' ₁With edge image M ₂Interpolation recover image M ' ₂, and with described M ₁And M ₂Each pixel is placed on M ' successively ₁And M ' ₂The position of odd-numbered line odd column is with M ' ₁And M ' ₂The pixel of all the other positions is as pixel to be inserted;

(4) to described interpolation recover image M ' ₂The known pixels of odd-numbered line odd column position averages, obtain this interpolation recover image M ' ₂The value of pixel to be inserted, the value of this pixel to be inserted is compared with predetermined threshold value T, if greater than T then this pixel to be inserted belongs to fringe region, otherwise belong to smooth region, threshold value T is between 180 to 220;

(5) described interpolation is recovered image M ' ₂Approximate as described interpolation recover image M ' ₁Edge image, with M ' ₁And M ' ₂The pixel of middle corresponding position is corresponding one by one, if M ' ₁In pixel to be inserted at smooth region, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the bicubic interpolation, obtains the value of this pixel to be inserted; If M ' ₁In pixel edge region to be inserted, then use M ' ₁The known pixels of odd-numbered line odd column position is carried out the autoregression interpolation, obtains the value of this pixel to be inserted, obtains the image M after the interpolation at last " ₁

2. image autoregression interpolation method according to claim 1, wherein described in the step (2) with image M behind edge detection operator Sobel and the down-sampling ₁The phase convolution, carry out as follows:

(2a) with the horizontal component of edge detection operator Sobel

With image M behind the down-sampling ₁The phase convolution;

(2b) with the vertical component of edge detection operator Sobel With image M behind the down-sampling ₁The phase convolution;

(2c) (2a) and the convolution results that (2b) obtains are made even root obtains image M behind Sobel operator and the down-sampling ₁Convolution results.

3. image autoregression interpolation method according to claim 1, wherein the use M ' described in the step (5) ₁The known pixels of odd-numbered line odd column position is carried out the autoregression interpolation, carries out as follows:

(5a) use interpolation recover image M ' ₁In be in the known pixels of 5 * 5 odd-numbered line odd columns of neighborhood of pixels to be inserted position of fringe region, to M ' ₁In the known pixels of 4 * 4 odd-numbered line odd columns of this neighborhood of pixels to be inserted position carry out optimal estimation, obtain M ' ₁In the weight vector of this pixel to be inserted:

(5b) according to weight vector

Utilize following formula calculate interpolation recover image M ' ₁In be in fringe region pixel y to be inserted ₅Value:

Wherein: y ₅Be the M ' that at every turn will insert ₁In be in the pixel to be inserted of fringe region, y ₁, y ₂, y ₃, y ₄, y ₆, y ₇, y ₈, y ₉Be M ' ₁Middle y ₅8 pixels to be inserted of neighborhood, y _iBe

In i pixel to be inserted, α _tBe

In t element, x _I-tBe M ' ₁Middle y _iThe known pixels of neighborhood t odd-numbered line odd column position, x _iBe M ' ₁Middle y ₅The known pixels of 4 odd-numbered line odd columns of neighborhood position, y _I-tBe M ' ₁Middle x _i4 pixels to be inserted of neighborhood;

(5c) will calculate interpolation recover image M ' ₁In be in fringe region pixel y to be inserted ₅Value, as interpolation recover image M ' ₁In be in the autoregression interpolation of fringe region pixel to be inserted.

Images